Target's SVP frames agent governance architecture as the real AI moat

At VB Transform 2026, Target senior vice president Siobhán Mc Feeney stated the company's AI edge does not come from the models it runs. Everything built around them does. "There's a lot in it. That to us is the moat," she said. "The models are great, and they're important. They're just not sufficient to be the competitive advantage."

The framing is direct: Mc Feeney described an agent governance architecture built around registration, certification, earned autonomy levels, and full lineage tracking from agent creation through runtime. This infrastructure, not model selection, is what Target positions as its operational moat. The claim is self-interested — Target is describing its own practices at a conference — but the specifics are concrete enough to evaluate.

One concrete example Mc Feeney cited involved a digital-twin simulation for men's shorts inventory across three Target stores in Long Beach during summer. The system predicted one store needed six to seven times more stock than the other two. Analysts initially questioned the result. The store sat less than two miles from the beach; the others were 10 to 12 miles inland. The recommendation was followed, and the stock sold through. Mc Feeney characterized results like this as the mechanism by which agentic systems earn higher autonomy over time.

Target structures agent autonomy as a four-level ladder. New agents begin by making observations without acting, then advance to suggesting actions while awaiting approval. The third level permits acting within defined guardrails. At the highest level Target currently operates, agents run end-to-end with a human in the loop. Mc Feeney stated that agents can lose earned autonomy if they drift from intended behavior, and that models that drift will be taken out of service.

The governance process begins before an agent is designed. Mc Feeney's team requires answers to a sequence of questions: what problem needs solving, whether an agent is the appropriate solution, what type of agent fits, and whether what is being called an agent is actually just a tool. When an agent is the right fit, builders must register and certify it. The registration process exists partly because existing solutions may already handle the problem, Mc Feeney said, and duplication is to be avoided.

Agent design then requires decisions about triggers — automation, engineer input, or scheduled runs — and about what the agent can access: which data, systems, tables, and databases. Observability follows. Mc Feeney said her team measures intended purpose, calibration, trajectory, runtime, and latency. The goal is full transparency that lets agents be adjusted over time. "You're talking about architecture and taxonomy and a data governance layer that absolutely had to be established," she said.

The autonomy-ladder model serves a specific function in Target's framing: it allows builders to move quickly without renegotiating guardrails on every task. Mc Feeney argued that when agents earn autonomy through demonstrated performance, builders gain a predictable operating envelope. "If you follow these guardrails, you follow security guidelines, you register the agent, and something still goes wrong, we have full lineage all the way through from the start," she said. "Our ability to recover is much better."

The cost-benefit angle also factors into model selection. Mc Feeney described frontier models as suited to complex tasks requiring large-scale data processing, such as heavy merchandising supply chains, while acknowledging they can be cost-prohibitive in other scenarios. "It's making sure there's always a cost benefit," she said.

What the source does not address is how Target validates that the autonomy ladder produces better outcomes than alternatives, whether other retailers are using comparable governance structures, or what the failure rate looks like when agents at higher autonomy levels underperform. The digital-twin example demonstrates one successful prediction but does not establish whether the governance architecture consistently delivers value relative to its overhead.

Mc Feeney also addressed workforce implications. Teams are operating at speeds that require new evaluation harnesses and continuous agent tracking. Builders and engineers managing human workers alongside AI systems need adapted skills. She described the career shift as "super exciting" and noted that builders now observe agents building while coaching humans who observe agents building — a layered dynamic requiring new forms of accountability.

The central claim — that governance infrastructure rather than model capability constitutes the moat — is framed as a competitive insight from a company describing its own practices. Whether the specific combination of registration, earned autonomy levels, and lineage tracking is genuinely differentiating or reflects sound operational discipline applied consistently is the question the source does not answer.